Skip to main content
used mini pcs

Local AI, measured

What a used mini PC can do with local LLMs

The right mini PC runs local models with Ollama, llama.cpp or LM Studio surprisingly well. Two numbers decide how it feels. How much RAM you have decides which models fit, and how fast that memory is decides how quickly they answer. We give measured speeds where they exist and say plainly where they don't.

What you need

SpecRecommendedWhy
CPU generationRyzen 7040/8040 (Phoenix/Hawk Point, Radeon 780M) for the best results; any AVX2-capable Intel 6th gen+ or Ryzen for small modelsPhoenix chips pair DDR5 with the Radeon 780M iGPU, which llama.cpp can use through Vulkan. Older chips run CPU-only. LM Studio requires AVX2 on x64, so check the exact CPU's spec page, especially on Pentium and Celeron units.
RAM32 GB minimum for useful models, 64 GB for 30B-class; always two matched sticksThe whole model file has to sit in memory, with room left for the context and the OS. Ollama's library lists Llama 3.1 8B at 4.9 GB, Qwen3 14B at 9.3 GB, Gemma 3 27B at 17 GB and Llama 3.1 70B (q4_K_M) at 43 GB.
Memory bandwidthDual-channel DDR5-5600 (about 89.6 GB/s theoretical) beats dual-channel DDR4-2666 (about 42.7 GB/s)Each generated token streams the whole active model through memory, so memory bandwidth sets the ceiling on token speed and extra cores don't raise it. Theoretical bandwidth is transfer rate × 8 bytes × channels, and a single-channel machine gets half.
Storage typeNVMe SSD, 1 TB if you keep several modelsModel files run from about 5 GB to over 40 GB each. NVMe loads a model into RAM quickly, and you'll do that often, because Ollama unloads idle models after 5 minutes by default.
NICsOne 1 GbE port is fine; 2.5 GbE is a bonusModels download once. After that, chat traffic from Open WebUI or an API client is tiny.
Idle wattsAbout 6 to 10 W for a Ryzen 7840HS mini at idleA Beelink SER7 (Ryzen 7 7840HS) measured 6 to 10 W idle, peaked around 77 to 79 W under load and settled to 44 to 46 W sustained. An LLM box idles most of the day, so idle draw dominates the yearly cost.
NoiseAudible under long generations; quiet at idleInference keeps the CPU or iGPU pinned for the length of every answer. The SER7 measured about 35 dBA under load in a 34 dBA room, but small chassis vary, so expect to hear the fan during long jobs.
iGPURadeon 780M via Vulkan is worth having; older Intel UHD iGPUs mean CPU-onlyllama.cpp's Vulkan backend has been measured on 780M machines (see below), and Ollama documents Vulkan as enabled by default when its backend is installed. Don't plan on ROCm, because the 780M isn't on Ollama's ROCm-supported Radeon list.

Which mini PCs fit

A model gets the “Local LLM” badge when it:

  • Intel 8th gen or newer, Intel N-series, or AMD Ryzen 2000 or newer (not Ryzen Embedded)
  • Takes at least 32 GB of RAM (manufacturer's stated maximum)
  • Has two or more memory slots, so RAM can run dual-channel (memory bandwidth sets LLM speed)
  • Has an M.2 slot that takes an NVMe SSD

We only use the makers’ specs. 33 of the 58 models we cover qualify.

Best overall

Minisforum UM790 Pro (Ryzen 9 7940HS)

Its chip has published numbers behind it. The UM790 Pro's Ryzen 9 7940HS with Radeon 780M is the exact chip in a llama.cpp Vulkan result of 19.91 tokens/sec on a 7B Q4_0 model. It takes dual-channel DDR5 SO-DIMMs up to 5600 MHz and 64 GB, with two M.2 2280 PCIe 4.0 slots and 2.5G Ethernet. Expect 8B to 14B models at conversational speed, and 30B-class models when you can wait.

On eBay now: A 7940HS for $325 if you have the parts. Budget for two DDR5-5600 SO-DIMMs and an NVMe SSD, and confirm the 120 W adapter is included. $325.00

View on eBay ↗: MinisForum UM790 PRO Mini PC w/ AMD Ryzen 9 7940HS Barebone SFF PC NO RAM/NO SSD

Best value

Lenovo ThinkCentre M75q Gen 2 Tiny

This is the cheap way into 30B-class models. It's a common off-lease 1-liter Lenovo rated for 64 GB of DDR4-3200 across two slots, with 6- and 8-core Ryzen PRO GE chips. It runs CPU-only and slower than the DDR5 machines, but 64 GB holds models such as Qwen3 30B (19 GB) that a 16 GB box can't load at all.

On eBay now: Linux Mint is already on it, so it's a ready Home Assistant or Docker box at $200 delivered. Ask what size the drive is. $200.00

View on eBay ↗: Lenovo ThinkCentre M75q Gen 2 Tiny Ryzen 3 Pro 4350GE | 16GB RAM | Linux Mint OS

What it costs to leave on all year

An always-on box spends almost all its time near idle, so idle draw sets the bill. Here’s the sum.

idle watts × 24 h × 365 days ÷ 1,000 = kWh per year · kWh × $0.1831/kWh = dollars per year

Electricity at 18.31¢/kWh, the U.S. average residential price, July 2026 (EIA). Your own rate is on your bill, so swap it into the sum. Real draw rises above idle whenever the box is working, and it depends on the drives, RAM, BIOS power settings and OS.

Memory bandwidth decides the speed

Large language models are mostly a memory problem. To produce each token, the machine has to read essentially every active weight in the model from RAM. A 5 GB model on a system that moves 40 GB per second hits a hard ceiling no matter how many cores you add. That's why a mini PC with dual-channel DDR5-5600 (about 89.6 GB/s theoretical) answers much faster than an older tiny with dual-channel DDR4-2666 (about 42.7 GB/s). A machine running a single stick, or an N100 with its single memory channel, is slower still. Always buy or install RAM as a matched pair.

Prompt processing, which means reading your question and documents before the answer starts, is the part that benefits from compute, so more cores or a GPU help there. Token generation, the part you watch scroll, is the part bandwidth limits.

What fits in 16, 32 and 64 GB

Ollama's README no longer publishes a RAM-per-model rule of thumb, so go by the file sizes in its model library and leave headroom for the OS, the context and anything else the box runs. Ollama's default context window is 4,096 tokens. Longer contexts and parallel requests need more memory on top of the model.

  • 16 GB runs 3B to 8B models comfortably. Llama 3.1 8B is 4.9 GB, Qwen3 8B is 5.2 GB and Gemma 3 4B is 3.3 GB. Gemma 3 12B (8.1 GB) fits but leaves little room for anything else.
  • 32 GB adds 14B models (Qwen3 14B, 9.3 GB) and the 27B to 32B class, with Gemma 3 27B at 17 GB, Qwen3 32B at 20 GB and Qwen3 30B at 19 GB.
  • 64 GB loads Llama 3.1 70B at q4_K_M (43 GB), but on a mini PC it's a batch-job model. It's very slow, so start it and go do something else.

One kind of model tips the math in a mini PC's favor. Qwen3 30B is a mixture-of-experts model, which Qwen describes as 30B-A3B, meaning only about 3B parameters are active for each token. You need RAM for the full 19 GB, but each token only reads a fraction of it, so MoE models are the most practical large models for bandwidth-limited machines.

Measured speeds on mini PC hardware

These are the only mini-PC-class numbers on this page, and each one comes from a published measurement.

  • On the Radeon 780M through llama.cpp's Vulkan backend, a Ryzen 9 7940HS hit 19.91 tokens/sec generation and 281.62 tokens/sec prompt processing on a 7B Q4_0 model (3.56 GiB). A Ryzen 7 8840HS posted 20.10 tokens/sec generation on the same test, and a Ryzen 7 7840U laptop with two DDR5-5600 sticks posted 18.22.
  • On an Intel N150, CPU-only through Ollama, Jeff Geerling's benchmarks measured a 16 GB GMKtec G3 Plus at 9.06 tokens/sec on Llama 3.2 3B and 2.13 tokens/sec on DeepSeek R1 14B.

A 780M machine running a 7B to 8B model feels conversational. An N-series box is pleasant with 3B models and painfully slow above that. We haven't found a published measurement for 8th to 10th-gen Intel tinies or DDR4 Ryzen machines. Our best read is that they're slow but usable for 7B to 8B chat and for overnight batch jobs like summarizing documents, and too slow for interactive use above about 14B.

Ollama, llama.cpp, LM Studio and Open WebUI

  • Ollama is the easiest server. One install and `ollama run <model>` get you a model plus an API other apps can call. By default it keeps a model in memory for 5 minutes after the last request. On AMD, its docs say Vulkan is enabled by default when the backend is installed. ROCm support is limited to listed discrete Radeon cards, and the 780M isn't one of them.
  • llama.cpp is the engine under many tools and the source of the Vulkan numbers above. Use it directly when you want to tune threads, context and GPU offload yourself.
  • LM Studio is a desktop app. On x64 it requires a CPU with AVX2 and recommends at least 16 GB of RAM, and on Linux it needs Ubuntu 20.04 or newer.
  • Open WebUI gives you a ChatGPT-style web interface for Ollama and OpenAI-compatible APIs. It runs entirely offline and installs as one Docker container.

When a mini PC is the wrong tool

A used mini PC won't run 30B to 70B models at conversational speed. There are two better routes for that. One is a desktop with a discrete GPU, where the model lives in fast VRAM. The other is an Apple Silicon Mac, whose unified memory is shared by CPU and GPU at much higher bandwidth. Apple lists the current Mac mini at 153 to 170 GB/s on the base chip and 307 GB/s on the M5 Pro, roughly 1.7 to 3.4 times the theoretical 89.6 GB/s of dual-channel DDR5-5600. A mini PC is the right tool when you want a cheap, quiet, always-on box for small models, private document summaries, or an endpoint for Home Assistant voice or n8n workflows.

If you mainly want to run cloud-model agents like Claude Code around the clock, you don't need local inference at all. See the AI agent server page. Shopping for a Ryzen mini? Beelink vs GMKtec vs Minisforum sorts out the brands.

Setup outline

  1. 1A BIOS supervisor (administrator) password locks boot order, UEFI and Secure Boot settings. Ask the seller if one is set before you pay, and skip the unit if they can't clear it.
  2. 2Check the BIOS security page for Computrace or Absolute on first power-up. Absolute's firmware module, found on machines from many PC makers, can reinstall its agent after you wipe the drive. If it's activated, get the seller to show the previous owner released it.
  3. 3Update the BIOS from the maker's support site, by serial number, before you tune memory. Out-of-date firmware is the usual reason an off-lease box acts strangely with RAM or NVMe drives.
  4. 4Install RAM as two matched SO-DIMMs, never a single stick. One stick halves memory bandwidth, which halves token speed. On DDR5 Ryzen boxes, confirm the speed in BIOS or with `dmidecode` after first boot.
  5. 5Run a full pass of MemTest86+. Large models fill RAM end to end, so a marginal stick that never bothered a desktop workload will crash inference.
  6. 6Install Ubuntu 24.04 LTS (or Windows 11 if you want LM Studio as a desktop app), then install Ollama and pull a small model first, such as an 8B, to confirm everything works.
  7. 7On a Radeon 780M machine, run `ollama ps` while a model is loaded. The processor column shows whether the model landed on the GPU, the CPU, or split across both.
  8. 8Add Open WebUI in Docker for a browser chat interface, and keep it on your LAN or behind Tailscale instead of exposing it to the internet.

Local LLM: common questions

How much RAM do I need for Ollama on a mini PC?

16 GB runs 3B to 8B models, and Llama 3.1 8B is 4.9 GB in Ollama's library. 32 GB adds 14B models and the 27B to 32B class, where Gemma 3 27B is 17 GB and Qwen3 32B is 20 GB. You need 64 GB for anything larger, such as Llama 3.1 70B at 43 GB, and it runs very slowly on a mini PC.

How fast is a Radeon 780M mini PC for local LLMs?

Published llama.cpp Vulkan results on 780M machines show about 18 to 20 tokens/sec of generation on a 7B Q4_0 model, with 19.91 on a Ryzen 9 7940HS and 20.10 on a Ryzen 7 8840HS. That's comfortably conversational for 7B to 8B models.

Can an Intel N100 or N150 run local AI?

Small models only. Jeff Geerling measured a 16 GB N150 mini PC at 9.06 tokens/sec on Llama 3.2 3B and 2.13 tokens/sec on DeepSeek R1 14B. Most N-series machines top out at 16 GB on a single memory channel.

Does Ollama support the Radeon 780M?

Not through ROCm, since Ollama's supported Radeon list covers discrete RX cards. Its docs say the Vulkan backend is enabled by default when installed, and llama.cpp's Vulkan backend has published 780M results. Run `ollama ps` to see whether a model loaded on the GPU or the CPU.

Mini PC or Mac mini for local LLMs?

For 30B-plus models at chat speed, a Mac with more unified memory is the stronger tool. Apple lists 153 to 307 GB/s of memory bandwidth on the current Mac mini, against about 90 GB/s theoretical for dual-channel DDR5-5600. A used x86 mini PC wins as a quiet, always-on Linux box running small models next to other services.

Is DDR4 OK for local LLMs?

It works at roughly half the bandwidth of dual-channel DDR5-5600, so expect noticeably slower generation. A 64 GB DDR4 box still beats a 16 GB DDR5 one where it counts, because it can load 30B-class models at all.

Sources

Run more on the same box

Upgrading? Sell the old one to us.

Tell us what you have (model, CPU, RAM, drive, whether there’s a BIOS password), add a few photos and your price, and we’ll reply with an offer by email. No listing, no fees.

Get an offer